[Mlir-commits] [mlir] [MLIR][LLVM] Add Continuous Loop Peeling transform to SCF (PR #71555)

[llvmlistbot at llvm.org]

================
@@ -105,6 +106,168 @@ static void specializeForLoopForUnrolling(ForOp op) {
   op.erase();
 }
 
+static LogicalResult splitLoopHelper(RewriterBase &b, scf::ForOp &forOp,
----------------
muneebkhan85 wrote:

Yes, the function can be called from peelForLoop. The test mlir/test/Dialect/SCF/for-loop-peeling.mlir passes when spliLoopHelper is called after the new splitBound has been computed as shown here

`static LogicalResult peelForLoop(RewriterBase &b, ForOp forOp,
                                 ForOp &partialIteration, Value &splitBound) {

  auto loc = forOp.getLoc();
  AffineExpr sym0, sym1, sym2;
  bindSymbols(b.getContext(), sym0, sym1, sym2);
  // New upper bound: %ub - (%ub - %lb) mod %step
  auto modMap = AffineMap::get(0, 3, {sym1 - ((sym1 - sym0) % sym2)});
  b.setInsertionPoint(forOp);
  splitBound = b.createOrFold<AffineApplyOp>(loc, modMap,
                                             ValueRange{forOp.getLowerBound(),
                                                        forOp.getUpperBound(),
                                                        forOp.getStep()});

  if(failed(splitLoopHelper(b, forOp, partialIteration, splitBound)))
    return failure();

  return success();
}`

But, should this change be a part of this patch? I think this is a separate change, as the current patch should only target the new addition of Continuous Loop Peeling?

https://github.com/llvm/llvm-project/pull/71555